EvilTokens showed what we forgot about 2013

Microsoft's disruption of EvilTokens — a platform that compromised 12,000 accounts through AI-assisted social engineering — frames this as a novel threat requiring urgent response. The real story is that we have known exactly how this succeeds for thirteen years and continue to act surprised each time the attack resurfaces in new clothes.

In 2013, attackers breached RSA SecureID and stole the master seeds that generate valid authentication tokens. The insight that emerged was brutal and simple: stealing a legitimate token beats cracking a password. Valid authentication factors, obtained through impersonation or coercion, bypass every technical safeguard designed to stop brute force. The lesson was absorbed by every serious security team and then immediately forgotten by everyone responsible for training users.

EvilTokens worked the same way. An AI chatbot impersonated IT support or company leadership, convinced users they needed to verify credentials urgently, collected valid tokens or login combinations. Weaponized them at scale. The platform succeeded not because AI made social engineering sophisticated. It made it cheap. A human impersonator can fool one person per conversation. A chatbot can run dozens simultaneously. The cone remains unchanged; only the channel gets cheaper each decade.

Organizations still frame the vulnerability as a problem of email formatting. They teach employees to spot misspellings, suspicious domains, awkward phrasing. This is security theater that mistakes the delivery mechanism for the actual flaw. The flaw is that humans respond to authority. A perfectly written request from someone claiming to be your manager triggers compliance faster than skepticism. An AI can write perfectly. It can also scale perfectly.

What changed between 2013 and now is not the human vulnerability. It is how many attackers can simultaneously exploit it. Every credential harvesting campaign succeeds on the same premise: the person receiving the message believes the person sending it has legitimate authority. Until security training teaches recognition of impersonation as a constant risk rather than recognition of suspicious formatting, each new technology will simply reduce the cost of running the same social engineering at larger scale. Microsoft stopped one instance. The vulnerability remains.